On Wed, Aug 10, 2016 at 2:10 PM, Heikki Linnakangas <hlinn...@iki.fi> wrote:
> Yeah, if the csnlog access turns out to be too expensive, we could do > something like this. In theory, you can always convert a CSN snapshot into > an old-style list of XIDs, by scanning the csnlog between the xmin and > xmax. That could be expensive if the distance between xmin and xmax is > large, of course. But as you said, you can have various hybrid forms, where > you use a list of XIDs of some range as a cache, for example. > > I'm hopeful that we can simply make the csnlog access fast enough, though. > Looking up an XID in a sorted array is O(log n), while looking up an XID in > the csnlog is O(1). That ignores all locking and different constant > factors, of course, but it's not a given that accessing the csnlog has to > be slower than a binary search of an XID array. FYI, I'm still fan of idea to rewrite XID with CSN in tuple in the same way we're writing hint bits now. I'm going to make prototype of this approach which would be enough for performance measurements. ------ Alexander Korotkov Postgres Professional: http://www.postgrespro.com The Russian Postgres Company
